Geekbench, long a standard fixture among benchmarking tools, has received its first major update in three years with the release of Geekbench 7. Developer Primate Labs has rebuilt the thinking behind its multi-core test from the ground up and refocused the GPU benchmark on machine learning and content creation. NVIDIA's CUDA returns to the list of supported APIs, and the reference machine used to calibrate scores has switched to a system built around AMD's Ryzen 7 7700.

A New Generation After Three Years, Still Free for Personal Use

Geekbench 7 was released on July 23, 2026. It arrives roughly three years after Geekbench 6 and is available simultaneously for Android, iOS, Windows, macOS, and Linux. Primate Labs has confirmed that the benchmark remains free for personal use, as it has been in previous versions.

Geekbench has earned its following by letting people compare everything from smartphones to workstations against a single yardstick. That makes any change to its workloads significant: the measuring stick itself has moved. The simplest way to read this update is as an effort to re-align that yardstick with how computers are actually used in 2026.

Rebuilding How the Multi-Core Score Is Calculated

The biggest change is in the multi-core test. Primate Labs argues that not every task in real applications is parallelized, and forcing parallel execution inside a benchmark distorts scores without saying anything useful about a device.

Geekbench 7 therefore runs a workload in multi-threaded mode only when the task it models actually runs multi-threaded in real applications. The clearest example given is the HTML5 Browser test. Because web browsers are single-threaded or only lightly threaded, that test is excluded from the multi-threaded suite. The goal is to break the pattern where a high core count translates directly into a high score, and instead reflect how quickly a device handles the work people really do.

New Media and Game Physics Workloads

On the CPU side, new media workloads measure audio and video encoding, decoding, and processing. They are modeled on video conferencing, screen sharing, and everyday content consumption, and there are 3 of them:

  • Encoding screen-sharing video with the AV1 codec, modeling the screen-sharing features of video conferencing apps
  • Compressing music and spoken-word audio with the Opus codec, modeling voice memo and podcast apps
  • Decoding video and audio while generating live captions with the Whisper speech recognition model, modeling playback with automatic subtitles enabled

Alongside these, a new Game Physics workload is built on Jolt Physics, the physics engine used in popular video games. The Photo Editor test has been expanded with a richer set of real-world edits, and the Photo Library workload now supports importing and processing relatively modern image formats such as JPEG XL and DNG.

The data being processed has grown heavier as well. The File Compression test now includes more varied archives spanning source code, object code, and text documents, while the PDF Viewer test covers everything from park maps to technical documents and academic papers. Developer and image processing tests use more assets in more formats, making the whole suite more demanding than Geekbench 6.

The GPU Benchmark Turns to Machine Learning and Creation, and CUDA Returns

The GPU benchmark has shifted its center of gravity toward the machine learning and content creation work that now defines GPU performance. On the machine learning side, it measures tracking faces in video and applying real-time filter effects, upscaling images with machine learning, and blurring backgrounds in video conferencing streams. Each mirrors a shipping feature: face filters in social apps, super resolution in creative tools, and virtual backgrounds in conferencing software.

For creation work, new image editing and synthesis workloads have been added, including RAW image processing, LUT-based video color grading, path tracing, and fluid simulation.

The notable change on the API side is the return of CUDA, which Primate Labs says came by popular demand. CUDA now sits alongside OpenCL, Vulkan, and Metal, so NVIDIA GPUs can be measured through the API that powers their most demanding workloads. With more people running GPUs for AI tasks, that should make the resulting numbers closer to real-world use.

Ryzen 7 7700 Becomes the Baseline, and Scores Cannot Be Compared Across Versions

The reference machine behind the scores has changed too. Geekbench 6 used Intel's Core i7-12700 as its baseline; Geekbench 7 calibrates a system with AMD's Ryzen 7 7700 to score 2,500 points. Because the baseline sits at 2,500, a score of 5,000 should be read as roughly twice the performance of the reference system.

Two caveats matter here. The baseline figure is a calibration value, not a score every Ryzen 7 7700 machine is expected to produce. And because the workloads and the calibration changed at the same time, Geekbench 6 and Geekbench 7 scores cannot be compared directly. Lining up numbers from older reviews or device databases invites the wrong conclusions, so cross-version comparisons are best avoided.

Geekbench 7 Pro Is 20 Percent Off Through August 6

Geekbench 7 itself is free for personal use, and Primate Labs is marking the launch with 20 percent off Geekbench 7 Pro through August 6. The single-user commercial license covers macOS, Windows, and Linux, and drops from 99 USD (about 16,000 yen) to 79 USD (about 13,000 yen). It is aimed at people managing multiple machines or running measurements as part of their job.

※1 USD = 163 JPY (as of July 25, 2026)

Summary

Geekbench 7 is a major update that rebuilds multi-core scoring around how real applications behave and adds media workloads using AV1, Opus, and Whisper along with a Jolt Physics-based Game Physics test. The GPU side leans into machine learning and content creation, and CUDA has returned to the supported APIs. The baseline has moved to a Ryzen 7 7700 system at 2,500 points, and Geekbench 6 scores are not directly comparable. With the yardstick for evaluating new devices shifting, it is worth getting into the habit of checking which version any published score was measured on.